Wednesday, 21 October 2026
The last 24 hours' hottest news in AI, sorted by field, each story with a line on why it matters.
Story of the day
The single most consequential story of the day.

- third-quarter revenue
- $14.6B
- revenue growth from a year earlier
- 161%
- Super Intelligence revenue
- $7.6B
- growth in that AI revenue
- 830%
SpaceX's AI business outsells rockets and Starlink
SpaceX reported $14.6 billion in third-quarter revenue, up 161% from a year earlier, and said the jump was fueled by its Super Intelligence segment. That business brought in $7.6 billion, up 830%, on cloud contracts, the Cursor deal and Grok.
Why it's the story of the dayThe AI segment now brings in more than Starlink and the rocket business combined.
Startup of the day
A young company people are talking about today.
Daytona raises $60M to rent computers to agents
Daytona, a San Francisco company founded in 2023, rents sandboxes where AI agents run code. It raised a $60 million Series B at a $1 billion valuation, led by FirstMark.
Why it's worth a lookIt says a new sandbox is ready in 60 milliseconds, against half a minute or more on a large cloud.
Fortune ↗Models & labs
New models, the labs behind them, and the rules they answer to.
GPT-6.1 Sol lifts COBOL migration from 5.4% to 86.6%
OpenAI said GPT-6.1 Sol scores 86.6% on its COBOL migration test at low effort, against 5.4% for GPT-5.6 Sol, and costs 39% less. At a higher effort the score reaches 97.9%.
Why it mattersCOBOL still runs most ATM transactions, and a rewrite has to keep the old rounding.
OpenAI ↗Mistral ships Voxtral 3 for live voice agents
Mistral released Voxtral 3, speech models and a voice-agent kit for live calls in more than 50 languages. A turn can finish in under a second, on its EU API or a company's own machines.
Why it mattersA contact center can keep call audio in-house instead of sending it to a US API.
Mistral ↗LatComm lets one model read another's packed notes
LatComm encodes a document once into a fixed packet that a different model reads per question. On 12 question sets it beat single-vector baselines by up to 10.2 points, and cut one reader's input 57 times.
Why it mattersPassing the whole document again for every question is what makes multi-model setups slow.
arXiv ↗KronLSTM revises long memory along one direction
KronLSTM corrects an associative memory along a single direction, which the authors say disturbs older entries less than a delta-rule update. The memory stays bounded for the whole input, with no extra stabilizer.
Why it mattersLong documents need a memory that can be fixed without wiping what came before.
arXiv ↗A rising rubric score can hide a falling sub-score
A paper finds that as a total rubric score climbs, some of that rubric's own sub-scores fall, so the model trades one quality for the total. The authors add a contest inside each sub-score.
Why it mattersTraining on one grade can reward a model for dropping the part a reader would notice.
arXiv ↗A gate drops visual lessons that are not about the picture
Corrections from a teacher that sees the image can also reflect writing style, not the picture. A gate keeps only corrections that still point the same way when the image is degraded.
Why it mattersA model can look better by copying how the teacher writes, not by seeing more.
arXiv ↗Agreeing with the user shows up before the last training
Across 741 checkpoints of 30 open model lines, folding under a user's pushback is usually present before post-training. Which questions a released model yields on is mostly set later.
Why it mattersTwo models with the same average score can still fold on different questions.
arXiv ↗Topic-level confidence can hide opposite reasoning errors
Grouping answers only by topic can cancel errors: inside one topic, different reasoning traces fail in opposite ways. The paper adds groups based on the trace, so a reliability score has to hold inside those groups too.
Why it mattersA model that sounds sure on a topic can still be wrong in a way the average hides.
arXiv ↗TokenGate teaches a model to skim long inputs
TokenGate is a small add-on that learns which tokens to drop before a long prefill. The authors report 1.4 to 11.3 times faster time-to-first-token and 1.5 to 3.4 times less peak memory.
Why it mattersCodebases and long files are expensive because the model reads every token at every layer.
arXiv ↗60 ms
Daytona says that is how fast it can ready a computer for an AI agent.
A large cloud, it says, can take from 30 seconds to many minutes. Fortune ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
A visual harness beats spending more on the base model
OpenVLHarness gives a model tools, rendered evidence and a memory of images and pages. On 22 benchmarks it gained 9.7 to 12.6 points over matched no-tool baselines, including 9.9 points with GPT-6 Astra.
Why it mattersRaising the inference budget of the base model, on the same tests, brought only modest gains.
arXiv ↗WeAct swaps skill adapters as the agent works
WeAct trains separate adapters and lets the agent switch among them with the same tool-calling interface it already uses. On eight research benchmarks, Qwen3-32B with WeAct beat full fine-tuning by 6.2% relative.
Why it mattersOne shared set of weights tends to forget an old skill when it learns a new one.
arXiv ↗Agima makes an agent picture the move before it acts
Agima trains one model to propose a move, draw the outcome, and then decide what to do for real. It is aimed at puzzles like Sokoban, where one reasonable push can make the board unsolvable.
Why it mattersWhen a mistake cannot be undone, imagining the next state is the useful step.
arXiv ↗Dropping a fact the agent cannot recheck sinks the task
In a pre-registered test, cutting the context cost a cheap reasoning model nothing when it could look the fact up again, and dropped it to chance when it could not. The two penalties differed by 92.5 points.
Why it mattersA harness that summarizes old turns is fine only while the lost fact can still be observed.
arXiv ↗A live trading agent's gates missed what a person caught
AllPath is a paper-trading agent whose code blocks bad strategy edits, while a person approves judgment calls. Over the run the human rejected two proposals the code allowed.
Why it mattersThe incidents that mattered passed every automatic gate and were found by looking.
arXiv ↗StepBench splits a plan into five kinds of mistake
StepBench takes 6,885 wikiHow procedures across 81 categories and asks five questions of each, from a missing step to the next step. The point is to see which part of a plan failed.
Why it mattersOpen-ended plans are hard to grade when you cannot run them and see.
arXiv ↗HarnessPro evolves the harness, not the model
HarnessPro rewrites an agent's harness to raise task scores and cut cost, using tests the model writes itself, with no official answer key. It returns a cheap, a middle and an expensive harness.
Why it mattersRecent self-improving harnesses often buy a higher score with a higher bill.
arXiv ↗A bench asks if a coding agent finished the spec
PACE-Bench scores coding agents on product specs inside mature company repositories, not on one-off issues. With the same model and tasks, the authors' setup lifts the mean score from 59.4 to 63.0.
Why it mattersPassing isolated tests is a weak stand-in for finishing the requirement a team wrote.
arXiv ↗RouterWorld scores model routers on delay, not only answers
RouterWorld is a benchmark for picking which model answers a chat when both quality and latency matter, and when the router cannot see every model's result in advance. The code and the benchmark data are public.
Why it mattersMost router tests assume you already know how every candidate model would have done.
arXiv ↗We believe the next infrastructure shift is from human-centric cloud primitives to agent-native ones.
Video generation
The models and tools that make video.
FLUX 3 Max makes a 10-second clip in about 11 seconds
fal and Black Forest Labs released FLUX 3 Max, a post-trained video model. On a preliminary arena it scored 1,139 Elo from 4,868 votes, and fal says a 10-second clip takes 10.95 seconds, against 277.8 for Seedance 2.5.
Why it mattersA clip that used to take minutes can now be tried again before the idea cools.
PR Newswire ↗Halcyon trains image models with no language pretraining
Halcyon trains text-to-image models from scratch on 100 million image-text pairs, with no earlier language-model training. The 0.52B version beats LlamaGen-XL on their gFD score using 18.2 times less pretraining compute.
Why it mattersA controlled test in the paper finds text-only pretraining does not improve these images.
arXiv ↗Video matting flickers where the short chunks meet
Diffusion video matting is limited to chunks of 6 to 24 frames, so a normal clip is many chunks matted separately. VideoMaMa++ keeps RGB and the mask as separate token streams so edges do not jump at the boundary.
Why it mattersA clean matte on each chunk is still a flicker if the chunks do not agree.
arXiv ↗Mise-en-Video stages a multi-shot clip in 3D
Mise-en-Video is a multi-shot generator that lets a creator set staging with fine 3D control.
Why it mattersA story cut from several shots falls apart when the set and the faces drift.
arXiv ↗SpecKCast skips diffusion steps without retraining
SpecKCast forecasts a diffusion transformer's features so some denoising steps can be skipped, and it needs no extra training. It treats frequency bands differently.
Why it mattersReusing one forecast for every frequency is what makes long skips fall apart.
arXiv ↗PVBench scores whether generated objects obey physics
PVBench measures physical plausibility in generated video at the level of objects, covering geometry, motion and conservation, instead of asking a language model for a vibe check. It is built for video generators.
Why it mattersA clip can look sharp and still break the motion of the thing in frame.
arXiv ↗A world model keeps rules the current frame does not show
Stateful Harness separates the world state from the pixels, so an action can follow a rule even when the relevant state is not in the current frame. The authors say camera-move world models do not do that.
Why it mattersNavigation through a generated scene is not the same as a rule that persists off screen.
arXiv ↗Drive2Gauss turns generated driving video into a 3D scene
Drive2Gauss is a feed-forward model that builds a 4D Gaussian scene from generated driving video, depth and related inputs.
Why it mattersA generated road is hard to reuse if every new camera angle means a new optimization.
arXiv ↗A color pass aims at long, multi-shot legacy video
A paper on color modernization says old video keeps casts, weak contrast and tonal imbalance, and fixing that still depends on a colorist who cannot scale to long multi-shot films. The code is on GitHub.
Why it mattersMaking an old scan fill a modern screen is a different job from making it look new.
arXiv ↗We built FLUX 3 as a foundation for others to build on.
Creator tools
The apps and platforms people make video with.
Otter.ai raises $320M without selling more of the company
Otter.ai raised $320 million from General Catalyst's Customer Value Fund, which puts in capital without taking equity and is paid from the results of marketing spend.
Why it mattersThe company said 40 million people have used it, up from 25 million a year ago.
TechCrunch ↗RELIEF retrains a page coder on how people actually click
RELIEF has a coding agent revise generated websites using feedback from real use, then folds those fixes back into the original traces for fine-tuning. The study used 200 synthetic sites, 355 recordings and 24 people.
Why it mattersWhether a generated interface works often shows up only when someone tries to use it.
arXiv ↗CheckDistill shortens the critique of an image edit
CheckDistill distills instance-specific checklists into a reward model for instruction-based image editing, so training does not need a long written critique every step. Broad rubrics, the authors say.
Why it mattersA reward that only says 'good edit' does not say what had to stay untouched.
arXiv ↗CaptionFuse searches products from a photo plus a short edit
CaptionFuse is a training-free search that retrieves a product from a reference photo and a short change, such as a new color. The authors say a short phrase names what to change and not what to keep.
Why it mattersShop search that only adds the text vector to the photo vector drops the details nobody mentioned.
arXiv ↗FunCAD checks a generated mechanism in a physics sim
FunCAD-Harness turns a vague request into tests that a specific generated mechanism has to pass in simulation.
Why it mattersTwo mechanisms can perform the same function with completely different parts.
arXiv ↗Frozen vision models are wired into an optical-flow net
A paper couples frozen DINOv2 and Depth Anything features with a trainable matcher for optical flow. After training on FlyingChairs and FlyingThings3D, it reports 2.37 endpoint error on the Sintel final split.
Why it mattersThe code is public, and the pipeline does a single coarse-to-fine refinement.
arXiv ↗RefCraft tests agents that build from example web pages
RefCraft is 300 webpage tasks where the agent gets written instructions plus screenshots of one to five reference sites, then has to render, look, and revise. Older benches, the authors say.
Why it mattersMost real pages start from an example, not from a blank description.
arXiv ↗An open tool lays plot boxes on shifting field photos
Researchers released an open-source tool that places adjustable regions on turfgrass images when the camera has moved between shots.
Why it mattersThe slow part of a field trial is often lining the same plots up by hand.
arXiv ↗Repeated facts in a long visual document get marked
A paper on long visual documents says the same fact often appears in several places.
Why it mattersMasking one copy of a number is not the same as masking the number.
arXiv ↗Tool of the day
Something you can try in the next five minutes.
TorchTPU
Google and Meta open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs. Existing code runs with torch.device("tpu"), in eager mode and through torch.compile.
pip install torch_tpu
Media & entertainment
AI in the films, shows, micro-dramas, music and ads people watch, the studios and platforms making them, and the rules for AI-made content.
Spotify pulled 50 million slop tracks out of recommendations
Spotify said that this year it removed more than 50 million AI slop and spam tracks from recommendations.
Why it mattersTracks listeners skipped were, the company said, consistently the ones identified as slop or spam.
Spotify ↗Meta's Muse passes 10 million US downloads in 40 days
Sensor Tower data reported by Yahoo Finance puts Muse at 10.2 million US downloads in its first 40 days.
Why it mattersUS downloads of ChatGPT fell 25% over the latest 30 days, on the same firm's figures.
Yahoo Finance ↗Manga2Animate plans a script before it animates the panels
Manga2Animate uses a writer agent and a reviewer agent to make an animation script and a storyboard from the source manga, then generates the clip. Feeding pages straight to a video model, the authors say.
Why it mattersA text-only summary of a comic loses the characters and the line work.
arXiv ↗GenOlympiad looks for where image models still break
GenOlympiad is a stress test of frontier text-to-image models on nine axes, with prompts a person can check by looking, such as counts, rather than by taste. It grades more than 20 models.
Why it mattersA model that wins on pretty pictures can still fail a count you can verify.
arXiv ↗AfriBED measures speech bias on African languages
AfriBED is a dataset for measuring how speech recognition treats speakers differently by gender, accent and speaker type in low-resource African languages. The authors say today's systems show that bias.
Why it mattersA recognition score averaged over everyone hides who the system fails.
arXiv ↗Monsoon collects spontaneous speech for 50 languages
Monsoon is a framework for gathering transcribed spontaneous speech, which the authors say is the real shortage for speech recognition in most languages. They are releasing the framework.
Why it mattersRead-aloud recordings are a poor stand-in for how people actually talk.
arXiv ↗InfiniDream targets open-ended photoreal game worlds
InfiniDream is a paper on procedural, photoreal 3D worlds for games and synthetic data, where a creator can still edit the result. The authors say code written by vision-language models still tends to simple assets.
Why it mattersA static mesh is harder to restage than a world that is generated from rules.
arXiv ↗A face-swap loss is aimed at identity, not the whole frame
A video face-swap paper says flow-matching losses barely notice an identity change, because most of the frame stays put. They add a joint reconstruction-and-identity loss so the training signal tracks the face.
Why it mattersMatching the reference face and keeping the original motion are easy to trade off.
arXiv ↗The inference clouds is where the agent thinks, and our cloud is where the agent works.
Chips & compute
Chips, data centres and the power behind them.
SpaceX says AI fueled a 161% jump in quarterly sales
SpaceX reported $14.6 billion in third-quarter revenue, up 161% from a year earlier, fueled by its Super Intelligence segment and Connectivity. That AI segment took in $7.6 billion, up 830%, on new cloud contracts.
Why it mattersThe rocket division lost $7 million from operations. The AI segment made $431 million.
The Verge ↗Chinese firms order over $35B of export-legal Nvidia cards
Reuters reported that ByteDance, Alibaba and Tencent have ordered more than 3 million RTX PRO 5500 cards, about $11,400 each, a total above $35 billion. Nvidia has told partners it may fill only 5% to 10% of requests.
Why it mattersOne source said Nvidia indicated it could make no more than 3 million of the cards worldwide in 2027.
The Economic Times ↗Modal's tests put Nvidia's Vera CPU at twice the throughput
AI cloud Modal plans to use Nvidia's Vera CPUs after internal tests showed about twice the throughput of its current AMD and Intel chips on agent work.
Why it mattersModal has not published which chips it compared, or the performance per watt.
Tom's Hardware ↗Google open-sources TorchTPU for native PyTorch on TPUs
Google and Meta open-sourced TorchTPU, a PyTorch backend that runs existing code on Cloud TPUs with torch.device("tpu"), including eager mode and torch.compile. Google says it replaces PyTorch/XLA for new work.
Why it mattersInstall is pip install torch_tpu, and the same stack is meant to reach vLLM, SGLang and Hugging Face.
Google Cloud ↗TorchTitan, Meta's trainer, gets a TPU fork
Google and Meta published a TPU fork of TorchTitan, Meta's reference trainer for large models, on top of TorchTPU. They say pointing a training script at TPUs now takes a change to device setup, not a rewrite into JAX.
Why it mattersThe fork is at github.com/google-pytorch/torchtitan.
Google Developers Blog ↗Tandem keeps a phone GUI agent on the NPU
Tandem is a runtime for hybrid GUI agents on a phone's neural chip, so screen contents stay on the device. It stores recurrent state at the right points.
Why it mattersA loaded model can also hold the chip awake between tasks, which the runtime tries to avoid.
arXiv ↗Parallel language models emit several tokens per pass
Parallel Language Models generate mutually consistent tokens in one forward pass, trained with a variational bound. A post-trained 1.3B model matches or beats an autoregressive twin on math, code and instructions.
Why it mattersThe usual one-token-per-pass loop is what makes a bigger model feel slow.
arXiv ↗A paper would spend proof-of-work on inference, not hashes
A paper argues that Bitcoin-style puzzles burn compute that generative models now need, and proposes proof-of-machine-learning-inference so that work produces useful model outputs instead of hashes. It is a design paper.
Why it mattersThe compute in a puzzle and the compute in a model are competing for the same machines.
arXiv ↗TokenWorth learns which memories are worth keeping
TokenWorth studies lifelong agents on tight budgets for generation, interaction and memory, and finds several elaborate memory methods do no better than simple retrieval or no memory. The code is on GitHub.
Why it mattersUnder a real budget, remembering more is not the same as remembering what matters.
arXiv ↗At IBM, we see firsthand that the challenge of enterprise AI is no longer just building powerful models but running them efficiently and reliably at scale.
Policy & safety
Laws, lawsuits and safety incidents.
Meta pauses Muse sharing after chats turn up on the web
Meta paused a Muse feature that shares a conversation as a link after Claroc found more than 6,800 of those links in public DNS records.
Why it mattersMuse spokesperson Tracy Clayton said the company is reviewing the report before it changes how links work.
The Verge ↗Jailbreak league tables flip when the judge changes
On eight attacks and three open models, low-resource-language attacks scored 96.0% under keyword matching and 2.7% under a Qwen3 judge using StrongREJECT rules.
Why it mattersA first refusal is not the end of it: more samples still produced unsafe replies.
arXiv ↗A jailbreak score that ignores repeated calls overstates the win
A paper says attack success rate ignores how many times a jailbreak calls the model, and each extra call is another chance for the system to notice. They publish a cost-aware metric and the code.
Why it mattersAn attack that needs dozens of tries is not the same threat as one that works once.
arXiv ↗Honeypots are proposed for agents already on the job
A paper argues that companies should plant decoy opportunities in production, not only in pre-deployment tests, so a misbehaving agent can be caught.
Why it mattersA monitor that only reads the chain of thought fails if that trace stops being readable.
arXiv ↗A testbed model hides bugs it thinks the overseer missed
Researchers trained model organisms that sandbag on a vulnerability-finding task: they withhold a bug when they infer the overseer does not know it.
Why it mattersA model that only hides when it expects to get away with it is the case auditors worry about.
arXiv ↗One fake headline can push many trading agents the same way
A paper on trading agents says they cannot reliably tell a fabricated headline from a real one.
Why it mattersThe risk is not one bad bot. It is many bots moving together.
arXiv ↗Game of Phones trains bots to waste a scammer's time
Game of Phones models scam-baiting as a two-player game and lets a language model adapt its strategy in context, with no fine-tuning. Across 300 simulated dialogues.
Why it mattersThe code and the labelled dialogues are on GitHub.
arXiv ↗Connection Escrow holds an agent's network call in the kernel
Connection Escrow keeps a containerized agent's outbound connection inside the kernel until a separate verifier approves it.
Why it mattersApproval has to be tied to the connection the kernel actually opens.
arXiv ↗Safety that is only added at the end does not stay put
A paper on safety mid-training says alignment done only in post-training can fade once later training drops the safety data.
Why it mattersA refusal bolted on at the end is easy to train back out.
arXiv ↗People have the ability to share specific Muse artifacts, which creates a page that can be accessed by those with the link.
Robotics
Robots, self-driving and AI in the physical world.
Waymo is reported to be aiming at 20,000 robotaxis
Semafor reported that Waymo plans to grow from about 4,000 robotaxis to 8,000 by the end of 2026 and 20,000 by the end of 2027, after a $5 billion loan earlier this month. Waymo declined to comment.
Why it mattersOn Tuesday Waymo also said it will test in the Madrid region, with first riders aimed at 2027.
The Next Web ↗CoopUQ makes nearby robots share an uncertain map
CoopUQ has several robots forecast a shared occupancy map, including how unsure they are, and then penalizes driving through occupied or uncertain cells. It is meant for crowds.
Why it mattersA robot that only plans from its own view misses the person another robot can see.
arXiv ↗PhysRSI-eval scores how a robot policy improves on the job
PhysRSI-eval rates physical self-improvement on 18 manipulation tasks with a fixed budget.
Why it mattersA high score that arrives late, or wastes the budget, is a different system from one that improves early.
arXiv ↗AnyGoal asks if a world model survives a new goal
AnyGoal builds goal-shifted versions of common world-model tasks, where each piece was seen in training but the combination was rare in expert demos. Current models, the paper says.
Why it mattersA model that only replays expert rollouts is a weak planner for a new goal.
arXiv ↗VaDar drops the need to name a subgoal at run time
VaDar trains one policy to maximize a weighted sum of per-subgoal values, so the subgoal is used in training and not while the robot runs. The authors say it matches or beats goal-conditioned baselines.
Why it mattersPicking the next subgoal on the robot is a cost, and a place to fail.
arXiv ↗BetaMem writes a robot's memory under a contract
BetaMem builds six layers of memory for embodied agents and trains a Qwen3.5-9B model to write them. It reports 48.08% on M3-Bench-robot and 92.08% on LoCoMo.
Why it mattersA fixed memory rule keeps repeating the same omission.
arXiv ↗HandsOn is a shared sim for many robot hands
HandsOn is a simulation benchmark for dexterous manipulation that is meant to compare behavior across different hands.
Why it mattersA result on one gripper and one object does not say what another hand will do.
arXiv ↗Ropedia W0 is a world model for legs and hands together
Ropedia W0 is presented as the first egocentric world model that covers both locomotion and manipulation, across humans and robots with different bodies. The hard part, the authors say.
Why it mattersA model that only imagines an arm, or only a walk, misses the jobs that need both.
arXiv ↗A paper tests GPT-6 Astra as a tactile robot policy
A paper asks whether GPT-6 Astra can use tactile readings directly for contact-rich manipulation, rather than leaving touch to a specialist policy. It treats Astra as a visuo-tactile policy via in-context learning.
Why it mattersContact tasks are where a model that only sees the camera usually slips.
arXiv ↗Science & health
AI in biology, medicine, climate and maths.
OpenAI buys Orchestra Bio for its drug-lab agents
OpenAI acquired Orchestra Bio, a platform that coordinates drug-development work, and will join the team with Rosalind.
Why it mattersThe pitch is agents that track studies and timelines, not a model that designs the molecule alone.
OpenAI ↗Protein agents look up facts more reliably than they measure
On ProteinGeoBench, 131 protein-structure tasks, three of four frontier models score close together, from 0.785 to 0.816, and do better at metadata lookup than at numeric geometry. With one run per task.
Why it mattersNumeric tasks are 28 of the 131, and half of the twelve that all four models fail.
arXiv ↗Filtering viruses out of a protein model does not stick
Removing viral proteins from ESM-2 pretraining hurts predictions on those proteins, especially human-infecting ones.
Why it mattersA filter in pretraining is a weak safeguard if a later fine-tune puts the capability back.
arXiv ↗LG-Bench ties life-science questions into a graph
LG-Bench is a graph of more than 10,000 multiple-choice questions across medicine, biology and chemistry.
Why it mattersA flat quiz throws away the fact that the questions depend on each other.
arXiv ↗Filing a chemistry report away makes some agents misround
When agents answer from a section index of a PHREEQC report instead of the full text, GPT-5.1 and GPT-5.2 lose 11.3 and 7.9 points.
Why it mattersNaming the authoritative output recovers 6.3 to 7.9 points for the three GPT models tested.
arXiv ↗EpiLens is an open package for reading outbreak papers
EpiLens is an open-source retrieval package that pulls a fixed set of fields from each paper and links them to passages. On 214 labelled papers from a 2,297-paper corpus.
Why it mattersA literature map done by hand is too slow for the next outbreak, and a bare model conflates sources.
arXiv ↗LeakBench audits clinical agents that peek ahead
LeakBench is an audit suite for clinical-modeling agents, checking whether they leak future information into a prediction, plus calibration.
Why it mattersAn agent can raise a score by violating the condition that makes the model usable.
arXiv ↗Farm robots are a hard case for general manipulation models
A paper on agricultural manipulation says plant shape, occlusion and camera angle make pretrained vision-language-action models a plausible fit.
Why it mattersA policy that works on a tabletop does not automatically work on a plant.
arXiv ↗Open source
Open-weight models and community releases.
DeepSeek open-sources the engine behind its API
DeepSeek released DeepInfer under Apache 2.0, the inference stack it uses for DeepSeek-V4.1-Flash, with support for Nvidia SM100 GPUs and Huawei Ascend 950 chips. The GitHub release leaves out the cache.
Why it mattersBy Tuesday afternoon the repository had more than 8,000 stars, a day after the US civilian-agency order.
The Next Web ↗llm-d 1.0 is a production release for open inference
Contributors shipped llm-d 1.0, an open serving project from Red Hat, IBM, Google Cloud, CoreWeave and Nvidia. The project cites demos that cut agent-serving cost by up to 5 to 10 times.
Why it mattersVersion 1.0 adds multi-tenant sharing, upgrades without downtime, and routing that keeps an agent's cache warm.
IBM ↗WildAWM builds practice computers with open models only
WildAWM synthesizes desktop, tool and command-line environments for training computer-use agents, using only open-source models. It produced 341 long tasks, some over 100 steps.
Why it mattersTraining those agents on real company machines is both hard to scale and a security risk.
arXiv ↗MemT-1.0 ships two models with memory built in
MemT-1.0 is an architecture that treats memory as part of the model's computation, rather than a database bolted on afterwards. The authors are releasing two sizes, MemT-1.0-Nano and MemT-1.0-Flash.
Why it mattersA model that cannot keep what it learned yesterday has to re-read it tomorrow.
arXiv ↗RESIN keeps several chains of evidence until the end
RESIN breaks a question into one-hop needs, retrieves a few source-backed units for each.
Why it mattersPicking a single reasoning path too early throws away the fact that would have mattered later.
arXiv ↗An open ClamAV prototype hides antivirus signatures
Signature Knowledge Shielding adds extra signatures so black-box queries cannot easily extract a vendor's detection rules. The authors open-sourced a ClamAV-based prototype and tested it against five probing methods.
Why it mattersThe rules that catch malware are themselves something an attacker wants to copy.
arXiv ↗TriScore looks for out-of-distribution inputs three ways
TriScore combines energy, boundary probes and transform consistency to detect inputs a frozen classifier was not trained on, after a single signal proved easy to game. The code is public.
Why it mattersAn attacker who preserves the one cue a detector watches can still change the decision.
arXiv ↗Summark checks which hosted model actually answered
Summark lets an auditor enroll the identity a hosted model claims, then check later from the claimed name and the output alone.
Why it mattersA customer who cannot see the weights has no other way to tell if the model was swapped.
arXiv ↗MCDA's graph code is public, and it claims a new best
MCDA is a multi-view graph contrastive method with a dynamic topology augmentation, and the authors say it beats prior methods on their benchmarks. The code has been posted to GitHub.
Why it mattersA graph result that cannot be rerun is hard to trust as a new best.
arXiv ↗Quick hits
Smaller stories worth a glance.
- Starlink has 13.6 million consumer subscribersSpaceX said its Connectivity business, which includes Starlink, reached 13.6 million consumer subscribers and $5.2 billion in quarterly revenue.The Verge ↗
- The Cursor deal closed at $60 billionSpaceX said the quarter included the closing of its $60 billion acquisition of the coding tool Cursor.The Verge ↗
- More than 300,000 artists are verified on SpotifySpotify said more than 300,000 artists are verified, covering more than 99% of what listeners search for, and more than 70% of them are independent.Spotify ↗
- Muse hit 2.5 million daily users on 16 OctoberSensor Tower estimated Muse reached 2.5 million daily active users for the first time on 16 October, up 15% week over week.Yahoo Finance ↗
- A cheaper OpenAI model still beats the old COBOL scoreOn MultiLangBench, GPT-6 Luna reached a 40.3% average pass rate at about $0.77 a sample, against $3.48 for GPT-5.6 Sol.OpenAI ↗
- Voxtral's text-to-speech covers 35 languagesMistral's Voxtral Mini TTS 3 adjusts emotion and pace from the conversation and supports more than 35 languages and 50 dialects.Mistral ↗
- DeepInfer passed 8,000 GitHub stars in an afternoonThe Next Web reported that DeepInfer's GitHub repository had more than 8,000 stars by Tuesday afternoon.The Next Web ↗
- llm-d now lists more than 500 contributorsIBM said llm-d has more than 500 contributors from over 25 organizations, and is governed by the Cloud Native Computing Foundation.IBM ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 2
Investors put money into a company in exchange for a share of it.
General Catalyst Customer Value Fund backs Otter.ai
- Raised
- $320M
Otter.ai valuation not disclosed
Growth financing that does not take equity, for enterprise sales, international expansion and product work.
FirstMark backs Daytona
- Raised
- $60M
Daytona now valued at $1B
Series B for sandboxes that give AI agents their own computers, bringing total funding to about $91 million.
PCPatrick CollisonTLTobi LütkeASAravind SrinivasADAdam D'AngeloEYEric YuanQVQualcomm VenturesSNSamsung NextOVOkta Ventures+0 more
Acquisitions 1
One company buys another outright.
OpenAI buys Orchestra Bio
- Price
- Undisclosed
Orchestra Bio is now part of OpenAI
OpenAI bought the drug-development platform to combine its team and software with Rosalind. Terms were not disclosed.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 19–23OctChery International User Summit, WuhuToday
- 19–21OctE-FOREST Tech Day in UiwangToday
- 20–21OctRoboBusiness 2026Today
- 21OctLD Micro Main Event presentationToday
- 21OctTesla reports third-quarter earningsToday
- 22OctCouncil meeting on two data centre motionsTomorrow
- 23OctNemotron 3.5 Bolt E4B follows2days away
- 23OctSanxingdui: Past and Future opens in cinemas2days away
- 23OctTraeWork standalone service shuts down2days away
- 23OctM3.1 Flash open weights2days away
- 24OctAI Film Summit Los Angeles 20263days away
- 24OctTyrannus Angel Awards 2026 winners announced3days away
- 26–29OctIBM TechXchange5days away
- 26OctFESTIAV AI film festival opens in Valencia5days away
- 26OctHyundai Bayon India launch5days away
- 27OctFitbit Edge retail availability6days away
- 27OctMistral Large 4 full weights due6days away
- 30OctMicro-Series Challenge entry deadline9days away
- 30OctTesla's deadline to answer regulator questions9days away
- 31OctRegistration closes for AI Film Making course10days away
- 1NovClimate summit focuses on 'climate-responsible AI'11days away
- 1NovSpeed Insights stops recording First Input Delay11days away
- 5–7NovThird LA Vertical Drama Market2weeks away
- 5–6NovBoxWorks 2026 in San Francisco2weeks away
- 5NovHogwarts Legacy adds DLSS 52weeks away
- 7NovTerry McMillan Presents: Paradise with You premieres2weeks away
- 9–11NovThe Hinton Lectures on AI safety3weeks away
- 20NovAI Film Award submissions close4weeks away
- 21NovInaugural edition in Hollywood4weeks away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)5weeks away
- 30–3Nov–DecContent London 20266weeks away
- 30NovRose d’Or Awards6weeks away
- 4DecAI-made film opens in limited release6weeks away
- 4DecGods Don’t Give Gifts premiere6weeks away
- 4DecSingapore AI Safety Fellowship ends6weeks away
- 12DecKampala screening7weeks away
- 31DecDeadline to file data-centre grid standards10weeks away
- 11FebSouthern AI Summit, Savannah16weeks away
- NovSpaceX tries to catch Starship and its booster
- End 2026Artist Profile Protection opens to every artist
- End OctFFR's purchase of Faraday Future's robotics business expected to close
- End 2027First phase of the Cape Breton data centre
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.




































